RCIM Paper Reference Benchmark

Purpose

This report turns the paper reference/RCIM_ML-compensation.pdf into a repository-owned benchmark package. Its role is to keep the paper findings easy to inspect during planning, training review, and colleague-facing status updates.

At the current repository state, the comparison is explicitly offline-only. The repository does not yet have the online compensation pipeline needed for a real Table 9 style comparison.

Guided Reading

Read the paper in this order:

1. Sections 2-3: problem setup, dataset structure, selected harmonics, and model evaluation logic.

2. Section 3.7 plus Tables 2-6: harmonic-level model ranking and the deployed model choices.

3. Sections 4-5 plus Table 9: TwinCAT/PLC integration and the final online compensation benchmark.

The most important paper message is not only that ML can fit the reducer TE, but that a deployable harmonic-wise prediction stack can be integrated into a PLC and produce large online reductions in TE during real motion profiles.

What The Paper Actually Says

Problem Setup

• The paper models rotational Transmission Error of an RV reducer as a function of input speed, applied torque, and oil temperature.

• The TE is reconstructed from selected harmonic components rather than from one monolithic end-to-end neural predictor.

• The paper explicitly separates:

  • offline harmonic prediction;

  • PLC integration;

  • online compensation validation.

Experimental Domain

• Total experimental samples: 1026

• Speed levels: 100 to 1800 rpm

• Torque levels: 0 to 1800 Nm

• Temperature levels: 25, 30, 35 C

Harmonic Structure

The paper emphasizes these harmonics as the practical basis for TE modeling and compensation:

• 0

• 1

• 3

• 39

• 40

• 78

• 81

• 156

• 162

• 240

The deployed compensation baseline then centers primarily on:

• 0

• 1

• 39

Additional validation variants also include:

• 40

• 78

Model Selection Summary

Paper-Level Conclusion

The paper does not end with a single universal winner across every harmonic. Instead, it converges toward a deployable stack dominated by tree and boosting models.

Harmonic-Level Selected Models

Harmonic	Selected Paper Model(s)
0	SVM
1	RF / LGBM
3	HGBM
39	HGBM
40	ERT / GBM
78	HGBM / RF
81	RF
156	ERT / RF
162	ERT
240	ERT

Why This Matters Here

• The paper is not arguing for a plain MLP-first deployment path.

• It favors models that are compact, interpretable enough for deployment work, and supported by the Beckhoff/TwinCAT stack.

• This is already directionally consistent with the current repository state, where the global offline winner is also tree-based.

Extracted Reference Metrics

Offline Prediction Validation

For unseen validation scenarios executed in TwinCAT-side prediction tests, the paper reports mean percentage errors along the TE function of:

• 2.6%

• 3.1%

• 4.7%

This is the most practical offline reference target that the repository can aim to reproduce before the online compensation loop exists.

Motion-Profile Test Conditions

Motion Profile	Max Speed [rpm]	Max Torque [Nm]	Temperature [C]
Robot	303	759	31.6
Cycloidal	500	370	26.7

Online Compensation Benchmark

Profile	Case	TE RMS [deg]	TE Max [deg]	Reduction [%]
Robot	No compensation	0.0478	0.0681	-
Robot	Comp (0,1,39)	0.0080	0.0325	83.3 / 52.4
Robot	Comp (0,1,39,40)	0.0078	0.0309	83.6 / 54.7
Robot	Comp (0,1,39,78)	0.0079	0.0319	83.5 / 53.2
Cycloidal	No compensation	0.0282	0.0534	-
Cycloidal	Comp (0,1,39)	0.0017	0.0044	94.0 / 91.7
Cycloidal	Comp (0,1,39,40)	0.0017	0.0062	94.0 / 88.3
Cycloidal	Comp (0,1,39,78)	0.0027	0.0020	90.5 / 96.3

Minimum Practical Targets For This Repository

Target A

Match or beat the paper on a comparable offline prediction benchmark.

Minimum repository target:

• reproduce a paper-comparable TE-curve validation protocol;

• reach <= 4.7% mean percentage error on unseen comparable scenarios.

Target B

Replicate the paper online compensation benchmark.

Minimum repository target:

• at least 83% robot-profile TE RMS reduction;

• at least 90% cycloidal-profile TE RMS reduction;

• and a cycloidal-profile TE max reduction in the same practical range as the paper benchmark.

Repository Comparison At The Current State

What Is Already Aligned

• The repository global offline winner is tree-based:

  • family: tree

  • model type: hist_gradient_boosting

  • run: te_hist_gbr_tabular

• The strongest current neural branch is still behind the tree winner:

  • family: residual_harmonic_mlp

• This is directionally consistent with the paper, where the deployable harmonic stack is dominated by tree and boosting models.

Comparison Structure To Preserve

The repository should now keep two explicit offline comparison tracks:

• Track 1: paper-faithful harmonic-wise benchmark

• Track 2: repository direct-TE comparable benchmark

The first track answers whether the repository can reproduce the paper’s own harmonic-wise logic. The second track answers whether the repository’s already trained direct-TE families can match or beat the paper at the level of final offline TE-curve prediction quality.

These two tracks must not be merged in reporting. Future paper-comparison tables should explicitly label each entry as either:

• paper-faithful harmonic-wise

• result-level comparable direct-TE

Canonical Track 1 Closure Rule

For Track 1, the primary closure criterion is no longer the best campaign winner under the shared offline evaluator.

The canonical Track 1 closure rule is now:

• reproduce the paper-facing cells in Tables 2, 3, 4, and 5;

• track status per harmonic target:

  • A_k MAE;

  • A_k RMSE;

  • phi_k MAE;

  • phi_k RMSE;

• track harmonic-level closure from Table 6;

• treat the harmonic-wise TE-curve evaluator only as supporting evidence.

Repository consequence:

• a harmonic-wise campaign winner may still be useful diagnostically;

• however, it does not close Track 1 by itself;

• Track 1 closes only when the canonical exact-paper table comparison shows the required paper-table cells as matched.

What Is Not Yet Comparable

• The repository now has a repository-owned harmonic-wise offline validation protocol, but the first baseline does not yet match the paper threshold.

• The repository does not yet have a harmonic-wise online compensation loop.

• The repository does not yet have TwinCAT-side or equivalent motion-profile compensation tests matching the paper’s Robot and Cycloidal profile benchmark.

• Therefore, the repository cannot yet claim a real comparison against the paper’s Table 9.

Primary Track 1 Status: Exact-Paper Table Replication

The primary Track 1 status must now be read from this canonical benchmark surface, with the older exact-paper validation report treated as historical supporting evidence:

• doc/reports/analysis/RCIM Paper Reference Benchmark.md

• doc/reports/analysis/validation_checks/2026-04-12-17-00-28_paper_reimplementation_rcim_exact_model_bank_rcim_exact_paper_model_bank_exact_paper_validation_tables_3_4_5_6_exact_paper_model_bank_report.md

Current exact-paper table-replication status from the canonical benchmark surface is:

• Table 2 amplitude MAE: 3/10 harmonics currently meet or beat the paper target;

• Table 3 amplitude RMSE: 6/10 harmonics currently meet or beat the paper target;

• Table 4 phase MAE: 5/9 harmonics currently meet or beat the paper target;

• Table 5 phase RMSE: 4/9 harmonics currently meet or beat the paper target;

• harmonic-level Table 6 closure: 1/10 fully matched, 8/10 partially matched, 1/10 not yet matched.

The highest-priority still-open harmonics now remain concentrated around:

• 1

• 3

• 81

• 162

• 240

Important interpretation:

• this exact-paper table status is the canonical Track 1 status;

• a harmonic-wise campaign result can inform which open cells to repair next;

• but it does not replace the table-level closure rule.

• the latest SVM open-cell repair campaign closed the Table 3 harmonic 0 gap on the canonical best-family surface and materially strengthened the SVM full-matrix row across Tables 2-5.

Deprecated Dashboard: Best-Envelope Reading

This dashboard is kept temporarily for historical continuity, but it is not the primary first-reading surface for Track 1.

Use the full paper-matrix replication dashboard below as the canonical view for model-by-model and harmonic-by-harmonic replication against the paper tables.

This section is now the canonical always-updated colleague-facing dashboard for the paper-facing Track 1 closure effort.

Maintenance rule:

• update this section after every material Track 1 progress step;

• keep the paper-side tables stable unless a source-reading correction is required;

• refresh the repository-side tables from the latest canonical exact-paper best run;

• after changing the full-matrix repository values, run scripts/reports/refresh_track1_benchmark_colored_markers.py so the colored 🟢/🟡/🔴 dashboard markers stay synchronized with the numeric cells;

• treat this section as open work until Track 1 reaches full closure.

Current repository evidence source for the dashboard:

• best current exact-paper run: exact_open_cell_paper_family_reference

• run instance id: 2026-04-13-22-08-40__exact_open_cell_paper_family_reference_campaign_run

• detailed supporting report: doc/reports/analysis/validation_checks/2026-04-13-22-09-00_paper_reimplementation_rcim_exact_model_bank_exact_open_cell_paper_family_reference_campaign_run_exact_paper_model_bank_report.md

Status legend used below:

• 🟢 target reached or beaten

• 🟡 not reached yet, but the positive gap is within 25% of the paper target and is therefore treated as near-target / acceptable follow-up

• 🔴 not reached and still materially open

Scope note:

• the paper-side tables below are repository-owned reconstructions of paper Tables 2-6;

• the repository-side tables are analogous tracking surfaces built from the current exact-paper validation outputs;

• Table 2 is necessarily a repository inference of the paper-facing deployed harmonic selection summary, because the repository needs one normalized view that can be compared directly against the current Track 1 best run.

Table 2 - Amplitude MAE

Paper-side repository-owned reconstruction:

Model	0	1	3	39	40	78	81	156	162	240
SVM	0.002600	5.60e-05	1.60e-04	1.50e-04	7.90e-05	2.60e-04	9.10e-05	4.40e-04	6.90e-04	2.90e-04
MLP	0.009500	0.006500	0.006500	0.005600	0.006900	0.007100	0.007400	0.006800	0.008100	0.005500
RF	0.003000	2.40e-05	2.00e-05	2.90e-05	2.60e-05	3.80e-05	1.10e-05	5.70e-05	6.80e-05	2.90e-05
DT	0.003400	2.90e-05	2.20e-05	4.00e-05	3.20e-05	5.90e-05	1.30e-05	6.30e-05	6.20e-05	5.10e-05
ET	0.003500	3.10e-05	2.40e-05	3.80e-05	3.20e-05	5.90e-05	1.80e-05	5.70e-05	8.80e-05	7.20e-05
ERT	0.003100	2.70e-05	2.30e-05	2.90e-05	2.30e-05	3.80e-05	1.20e-05	1.70e-05	2.30e-05	2.40e-05
GBM	0.003100	2.70e-05	2.10e-05	2.80e-05	2.70e-05	3.90e-05	1.20e-05	6.10e-05	7.10e-05	3.00e-05
HGBM	0.002400	2.70e-05	1.50e-05	2.10e-05	2.60e-05	2.70e-05	1.20e-05	1.00e-04	1.70e-04	3.50e-05
XGBM	0.002500	5.50e-05	8.10e-05	1.10e-04	6.60e-05	1.10e-04	4.60e-05	2.30e-04	2.60e-04	1.40e-04
LGBM	0.002500	2.70e-05	1.80e-05	2.40e-05	2.70e-05	3.00e-05	1.20e-05	9.00e-05	1.60e-04	3.20e-05

Repository-side analogous Track 1 table:

Harmonic	Paper Best Family	Paper Target MAE	Repo Best Family	Repo Best MAE	Gap Vs Paper	Status
0	HGBM	0.002400	XGBM	0.002465	6.46e-05	🟡
1	RF	2.40e-05	HGBM	2.54e-05	1.45e-06	🟡
3	HGBM	1.50e-05	HGBM	1.82e-05	3.21e-06	🟡
39	HGBM	2.10e-05	HGBM	2.34e-05	2.35e-06	🟡
40	ERT	2.30e-05	RF	2.21e-05	-9.06e-07	🟢
78	HGBM	2.70e-05	LGBM	2.46e-05	-2.41e-06	🟢
81	RF	1.10e-05	ERT	1.09e-05	-1.26e-07	🟢
156	ERT	1.70e-05	ERT	3.47e-05	1.77e-05	🔴
162	ERT	2.30e-05	ERT	4.49e-05	2.19e-05	🔴
240	ERT	2.40e-05	GBM	3.38e-05	9.84e-06	🔴

Quick read for Table 2:

• amplitude MAE is strongest on 40, 78, and 81;

• the near-closure amplitude MAE columns are 0, 1, 3, and 39;

• the dominant unresolved amplitude MAE columns are 156, 162, and 240.

Table 3 - Amplitude RMSE

Paper-side repository-owned reconstruction:

Model	0	1	3	39	40	78	81	156	162	240
SVM	0.003300	7.40e-05	1.80e-04	1.80e-04	9.50e-05	3.30e-04	1.00e-04	8.80e-04	0.002200	4.70e-04
MLP	0.0140	0.0120	0.0120	0.0100	0.0140	0.0130	0.0150	0.0130	0.0160	0.0100
RF	0.004100	3.50e-05	3.00e-05	3.80e-05	3.70e-05	5.60e-05	1.50e-05	1.70e-04	2.20e-04	5.40e-05
DT	0.004900	4.00e-05	3.30e-05	5.30e-05	4.50e-05	8.20e-05	1.80e-05	2.00e-04	1.70e-04	1.10e-04
ET	0.004500	4.20e-05	3.50e-05	5.10e-05	4.30e-05	8.50e-05	2.70e-05	1.90e-04	3.80e-04	1.80e-04
ERT	0.004000	3.70e-05	3.40e-05	4.00e-05	3.60e-05	5.70e-05	1.60e-05	1.30e-04	1.60e-04	4.20e-05
GBM	0.004000	3.60e-05	3.10e-05	3.90e-05	3.90e-05	5.50e-05	1.60e-05	1.70e-04	2.20e-04	4.70e-05
HGBM	0.003400	3.60e-05	2.50e-05	3.20e-05	3.80e-05	4.50e-05	1.60e-05	2.50e-04	5.00e-04	7.40e-05
XGBM	0.003500	7.10e-05	1.00e-04	1.30e-04	8.70e-05	1.50e-04	6.00e-05	5.40e-04	7.50e-04	2.10e-04
LGBM	0.003500	3.70e-05	2.60e-05	3.30e-05	3.80e-05	4.60e-05	1.60e-05	2.20e-04	4.70e-04	6.20e-05

Repository-side analogous Track 1 table:

Harmonic	Paper Best Family	Paper Target RMSE	Repo Best Family	Repo Best RMSE	Gap Vs Paper	Status
0	SVM	0.003300	SVM	0.003110	-1.90e-04	🟢
1	RF	3.50e-05	HGBM	3.52e-05	1.76e-07	🟢
3	HGBM	2.50e-05	HGBM	2.57e-05	7.42e-07	🟡
39	HGBM	3.20e-05	HGBM	3.17e-05	-3.47e-07	🟢
40	ERT	3.60e-05	RF	3.28e-05	-3.24e-06	🟢
78	HGBM	4.50e-05	LGBM	3.57e-05	-9.30e-06	🟢
81	RF	1.50e-05	RF	1.82e-05	3.22e-06	🟡
156	ERT	1.30e-04	ERT	1.05e-04	-2.46e-05	🟢
162	ERT	1.60e-04	ERT	1.44e-04	-1.64e-05	🟢
240	ERT	4.20e-05	RF	5.47e-05	1.27e-05	🔴

Table 4 - Phase MAE

Paper-side repository-owned reconstruction:

Model	1	3	39	40	78	81	156	162	240
SVM	0.002200	0.0330	0.0270	0.0610	0.1900	0.1300	1.200	0.4900	0.4900
MLP	0.007200	0.0650	0.0620	0.0800	0.1600	0.1500	1.900	0.7800	0.7000
RF	0.002000	0.0240	0.0280	0.0370	0.0740	0.0530	0.5100	0.2300	0.2500
DT	0.002100	0.0300	0.0360	0.0430	0.0900	0.0660	0.5200	0.2000	0.2300
ET	0.002400	0.0310	0.0350	0.0510	0.0940	0.0870	0.7100	0.2800	0.2600
ERT	0.002200	0.0270	0.0280	0.0400	0.0760	0.0560	0.5300	0.2000	0.2300
GBM	0.002000	0.0240	0.0300	0.0360	0.0740	0.0530	0.5400	0.2500	0.2900
HGBM	0.001900	0.0200	0.0210	0.0400	0.0910	0.0570	0.7400	0.3500	0.3600
XGBM	0.001900	0.0240	0.0320	0.0610	0.1400	0.0910	0.9600	0.5400	0.3900
LGBM	0.001800	0.0210	0.0210	0.0400	0.0950	0.0550	0.7400	0.3500	0.3400

Repository-side analogous Track 1 table:

Harmonic	Paper Best Family	Paper Target MAE	Repo Best Family	Repo Best MAE	Gap Vs Paper	Status
1	LGBM	0.001800	HGBM	0.001846	4.64e-05	🟡
3	HGBM	0.0200	GBM	0.0238	0.003757	🟡
39	HGBM	0.0210	LGBM	0.0204	-6.25e-04	🟢
40	GBM	0.0360	ERT	0.0345	-0.001478	🟢
78	GBM	0.0740	RF	0.0516	-0.0224	🟢
81	GBM	0.0530	LGBM	0.0475	-0.005526	🟢
156	RF	0.5100	ERT	0.3967	-0.1133	🟢
162	DT	0.2000	ERT	0.2125	0.0125	🟡
240	DT	0.2300	ERT	0.2699	0.0399	🟡

Table 5 - Phase RMSE

Paper-side repository-owned reconstruction:

Model	1	3	39	40	78	81	156	162	240
SVM	0.003100	0.0420	0.0440	0.0970	0.3200	0.2000	1.800	1.100	1.100
MLP	0.0130	0.0840	0.0770	0.1100	0.2400	0.2200	2.200	1.200	1.100
RF	0.002800	0.0330	0.0430	0.0550	0.1600	0.0820	1.200	0.6800	0.6300
DT	0.002800	0.0420	0.0610	0.0610	0.2000	0.1000	1.300	0.7300	0.6700
ET	0.003300	0.0460	0.0620	0.0740	0.2300	0.1500	1.500	0.9300	0.6800
ERT	0.003600	0.0400	0.0440	0.0600	0.1800	0.1100	1.200	0.6400	0.5800
GBM	0.002600	0.0340	0.0450	0.0550	0.1800	0.0840	1.300	0.7100	0.7100
HGBM	0.002500	0.0290	0.0270	0.0600	0.1900	0.0850	1.300	0.7000	0.7400
XGBM	0.002800	0.0330	0.0430	0.0890	0.2300	0.1300	1.400	0.8100	0.7600
LGBM	0.002500	0.0300	0.0280	0.0600	0.1900	0.0820	1.300	0.7000	0.7100

Repository-side analogous Track 1 table:

Harmonic	Paper Best Family	Paper Target RMSE	Repo Best Family	Repo Best RMSE	Gap Vs Paper	Status
1	HGBM	0.002500	GBM	0.002510	1.01e-05	🟡
3	HGBM	0.0290	HGBM	0.0343	0.005302	🟡
39	HGBM	0.0270	HGBM	0.0326	0.005648	🟡
40	GBM	0.0550	ERT	0.0541	-8.81e-04	🟢
78	RF	0.1600	RF	0.1250	-0.0350	🟢
81	LGBM	0.0820	RF	0.0681	-0.0139	🟢
156	ERT	1.200	ERT	0.9129	-0.2871	🟢
162	ERT	0.6400	ERT	0.7186	0.0786	🟡
240	ERT	0.5800	ERT	0.7573	0.1773	🔴

Table 6 - Harmonic Closure Summary

Paper-side repository-owned reconstruction:

k	Paper A*_k	Paper phi*_k
0	SVM	-
1	RF	LGBM
3	HGBM	HGBM
39	HGBM	HGBM
40	ERT	GBM
78	HGBM	RF
81	RF	RF
156	ERT	RF
162	ERT	ERT
240	ERT	ERT

Repository-side analogous Track 1 table:

k	Paper A*_k	Repo Best Ampl Family	Ampl Status	Paper phi*_k	Repo Best Phase MAE Family	Repo Best Phase RMSE Family	Phase MAE Status	Phase RMSE Status	Harmonic Status	Overall Color
0	SVM	SVM	met_paper_target	-	ERT	ERT	not_applicable	not_applicable	fully_matched_tables_3_6	🟢
1	RF	HGBM	above_paper_target	LGBM	HGBM	GBM	above_paper_target	above_paper_target	not_yet_matched_tables_3_6	🔴
3	HGBM	HGBM	above_paper_target	HGBM	GBM	HGBM	above_paper_target	above_paper_target	partially_matched_tables_3_6	🟡
39	HGBM	HGBM	met_paper_target	HGBM	LGBM	HGBM	met_paper_target	above_paper_target	partially_matched_tables_3_6	🟡
40	ERT	RF	met_paper_target	GBM	ERT	ERT	met_paper_target	met_paper_target	partially_matched_tables_3_6	🟡
78	HGBM	LGBM	met_paper_target	RF	RF	RF	met_paper_target	met_paper_target	partially_matched_tables_3_6	🟡
81	RF	RF	above_paper_target	RF	LGBM	RF	met_paper_target	met_paper_target	partially_matched_tables_3_6	🟡
156	ERT	ERT	met_paper_target	RF	ERT	ERT	met_paper_target	met_paper_target	partially_matched_tables_3_6	🟡
162	ERT	ERT	met_paper_target	ERT	ERT	ERT	above_paper_target	above_paper_target	partially_matched_tables_3_6	🟡
240	ERT	RF	above_paper_target	ERT	ERT	ERT	above_paper_target	above_paper_target	partially_matched_tables_3_6	🟡

Current dashboard reading:

• fully green harmonics: none yet

• partial yellow harmonics: 3, 39, 40, 78, 81, 156, 162, 240

• fully red harmonics: 0, 1

This means Track 1 is closer to structural closure than before, but still blocked by:

• the low-order amplitude and phase pair at 1;

• the structural amplitude mismatch at 0;

• the late-phase numeric gaps at 162 and 240;

• and the still-open phase-side gap at 3.

Canonical Track 1 Dashboard: Full Paper-Matrix Replication

This dashboard is now the canonical first-reading surface for the clarified first objective of Track 1:

• reproduce the paper matrices family by family;

• keep the exact paper model rows intact;

• read campaign progress as row replication, not only as best-envelope closure.

Current repository evidence source for the full matrices:

• latest full-matrix row-reproduction campaign: track1_full_matrix_family_reproduction_campaign_2026_04_14_13_50_51

• execution window: 2026-04-14 14:12:14+02:00 to 2026-04-14 14:15:18+02:00

• supporting campaign report: doc/reports/campaign_results/2026-04-14-14-35-29_track1_full_matrix_family_reproduction_campaign_results_report.md

Status legend used in the repository matrices:

• 🟢 repository value reached or beat the paper cell

• 🟡 repository value is still above the paper cell, but the positive gap is within 25% of the paper value

• 🔴 repository value is still materially above the paper cell

Important scope boundary:

• the matrices below use the exact same model-family rows as the paper: SVM, MLP, RF, DT, ET, ERT, GBM, HGBM, XGBM, LGBM

• the repository values come from our repository-owned implementations of those model families under the exact-paper validation branch

• Tables 2 and 6 remain useful summary/context surfaces, but the primary first Track 1 replication target is the full matrix structure of Tables 3, 4, and 5

Table 2 - Amplitude MAE Full-Matrix Replication

Paper-side repository-owned reconstruction:

Model	0	1	3	39	40	78	81	156	162	240
SVM	0.002600	5.60e-05	1.60e-04	1.50e-04	7.90e-05	2.60e-04	9.10e-05	4.40e-04	6.90e-04	2.90e-04
MLP	0.009500	0.006500	0.006500	0.005600	0.006900	0.007100	0.007400	0.006800	0.008100	0.005500
RF	0.003000	2.40e-05	2.00e-05	2.90e-05	2.60e-05	3.80e-05	1.10e-05	5.70e-05	6.80e-05	2.90e-05
DT	0.003400	2.90e-05	2.20e-05	4.00e-05	3.20e-05	5.90e-05	1.30e-05	6.30e-05	6.20e-05	5.10e-05
ET	0.003500	3.10e-05	2.40e-05	3.80e-05	3.20e-05	5.90e-05	1.80e-05	5.70e-05	8.80e-05	7.20e-05
ERT	0.003100	2.70e-05	2.30e-05	2.90e-05	2.30e-05	3.80e-05	1.20e-05	1.70e-05	2.30e-05	2.40e-05
GBM	0.003100	2.70e-05	2.10e-05	2.80e-05	2.70e-05	3.90e-05	1.20e-05	6.10e-05	7.10e-05	3.00e-05
HGBM	0.002400	2.70e-05	1.50e-05	2.10e-05	2.60e-05	2.70e-05	1.20e-05	1.00e-04	1.70e-04	3.50e-05
XGBM	0.002500	5.50e-05	8.10e-05	1.10e-04	6.60e-05	1.10e-04	4.60e-05	2.30e-04	2.60e-04	1.40e-04
LGBM	0.002500	2.70e-05	1.80e-05	2.40e-05	2.70e-05	3.00e-05	1.20e-05	9.00e-05	1.60e-04	3.20e-05

Repository-side analogous matrix:

Model	0	1	3	39	40	78	81	156	162	240
SVM	🟢 0.002503	🟢 5.31e-05	🟢 1.57e-04	🟢 1.49e-04	🟡 8.20e-05	🟢 2.52e-04	🟢 8.23e-05	🟡 4.61e-04	🟢 6.82e-04	🟢 2.52e-04
MLP	🟡 0.0106	🔴 0.008904	🔴 0.009453	🔴 0.007449	🟡 0.006993	🟢 0.007089	🟡 0.00879	🟡 0.007806	🟢 0.007446	🔴 0.008946
RF	🟡 0.003114	🟡 2.65e-05	🟢 1.89e-05	🟢 2.73e-05	🟢 2.21e-05	🟢 3.72e-05	🟢 1.10e-05	🟢 5.31e-05	🟢 5.28e-05	🟡 3.40e-05
DT	🟡 0.00351	🟡 3.10e-05	🟡 2.32e-05	🟢 3.66e-05	🟢 2.97e-05	🟢 5.56e-05	🟡 1.52e-05	🔴 8.92e-05	🟡 6.70e-05	🟢 4.51e-05
ET	🟢 0.003385	🟢 3.02e-05	🟢 2.38e-05	🟡 4.74e-05	🟢 2.59e-05	🟡 6.03e-05	🟢 1.31e-05	🔴 7.83e-05	🟢 7.69e-05	🟡 8.09e-05
ERT	🟡 0.003229	🟢 2.63e-05	🟢 2.11e-05	🟢 2.66e-05	🟢 2.27e-05	🟢 3.56e-05	🟢 1.09e-05	🔴 3.47e-05	🔴 4.49e-05	🔴 3.81e-05
GBM	🟡 0.003237	🟡 2.71e-05	🟢 1.88e-05	🟢 2.68e-05	🟢 2.55e-05	🟡 3.92e-05	🟢 1.15e-05	🟡 6.34e-05	🟡 7.24e-05	🟡 3.38e-05
HGBM	🟡 0.002505	🟢 2.54e-05	🟡 1.82e-05	🟡 2.34e-05	🟢 2.48e-05	🟢 2.54e-05	🟢 1.16e-05	🟡 1.01e-04	🟢 1.38e-04	🟡 3.85e-05
XGBM	🟢 0.002465	🟢 5.29e-05	🟢 7.21e-05	🟢 9.31e-05	🟢 6.44e-05	🟡 1.17e-04	🟡 4.66e-05	🟡 2.74e-04	🟢 2.21e-04	🔴 1.87e-04
LGBM	🟡 0.002613	🟢 2.67e-05	🟡 1.89e-05	🟢 2.35e-05	🟢 2.53e-05	🟢 2.46e-05	🟡 1.26e-05	🟡 1.05e-04	🟢 1.21e-04	🟡 3.41e-05

Quick read for Table 2:

• amplitude MAE is strongest row-wise on RF, HGBM, XGBM, and LGBM;

• the hardest amplitude MAE columns remain 156, 162, and 240;

• 40, 78, and 81 are the healthiest amplitude MAE columns.

Table 3 - Amplitude RMSE Full-Matrix Replication

Paper-side repository-owned reconstruction:

Model	0	1	3	39	40	78	81	156	162	240
SVM	0.003300	7.40e-05	1.80e-04	1.80e-04	9.50e-05	3.30e-04	1.00e-04	8.80e-04	0.002200	4.70e-04
MLP	0.0140	0.0120	0.0120	0.0100	0.0140	0.0130	0.0150	0.0130	0.0160	0.0100
RF	0.004100	3.50e-05	3.00e-05	3.80e-05	3.70e-05	5.60e-05	1.50e-05	1.70e-04	2.20e-04	5.40e-05
DT	0.004900	4.00e-05	3.30e-05	5.30e-05	4.50e-05	8.20e-05	1.80e-05	2.00e-04	1.70e-04	1.10e-04
ET	0.004500	4.20e-05	3.50e-05	5.10e-05	4.30e-05	8.50e-05	2.70e-05	1.90e-04	3.80e-04	1.80e-04
ERT	0.004000	3.70e-05	3.40e-05	4.00e-05	3.60e-05	5.70e-05	1.60e-05	1.30e-04	1.60e-04	4.20e-05
GBM	0.004000	3.60e-05	3.10e-05	3.90e-05	3.90e-05	5.50e-05	1.60e-05	1.70e-04	2.20e-04	4.70e-05
HGBM	0.003400	3.60e-05	2.50e-05	3.20e-05	3.80e-05	4.50e-05	1.60e-05	2.50e-04	5.00e-04	7.40e-05
XGBM	0.003500	7.10e-05	1.00e-04	1.30e-04	8.70e-05	1.50e-04	6.00e-05	5.40e-04	7.50e-04	2.10e-04
LGBM	0.003500	3.70e-05	2.60e-05	3.30e-05	3.80e-05	4.60e-05	1.60e-05	2.20e-04	4.70e-04	6.20e-05

Repository-side analogous matrix:

Model	0	1	3	39	40	78	81	156	162	240
SVM	🟢 0.003110	🟢 7.01e-05	🟢 1.79e-04	🟢 1.78e-04	🟡 9.63e-05	🟢 3.15e-04	🟢 9.39e-05	🟡 0.001048	🟢 0.002181	🟡 4.86e-04
MLP	🟡 0.0141	🔴 0.0156	🔴 0.0155	🔴 0.0126	🟢 0.0121	🟢 0.0120	🟢 0.0150	🟢 0.0127	🟢 0.0130	🔴 0.0165
RF	🟡 0.004114	🟡 3.72e-05	🟢 2.73e-05	🟢 3.74e-05	🟢 3.28e-05	🟡 5.67e-05	🟡 1.82e-05	🟡 1.97e-04	🟢 1.52e-04	🟡 5.47e-05
DT	🟢 0.004879	🟡 4.31e-05	🟡 3.34e-05	🟢 5.06e-05	🟡 4.51e-05	🟢 7.97e-05	🔴 2.38e-05	🔴 2.98e-04	🟡 1.97e-04	🟢 7.18e-05
ET	🟢 0.004280	🟢 4.15e-05	🟡 3.54e-05	🔴 7.57e-05	🟢 3.77e-05	🟡 8.65e-05	🟢 2.01e-05	🔴 2.65e-04	🟢 2.86e-04	🔴 2.74e-04
ERT	🟡 0.004201	🟡 3.75e-05	🟢 3.17e-05	🟢 3.80e-05	🟢 3.34e-05	🟢 5.21e-05	🟡 1.85e-05	🟢 1.05e-04	🟢 1.44e-04	🔴 7.22e-05
GBM	🟡 0.004261	🟡 3.79e-05	🟢 2.66e-05	🟢 3.80e-05	🟢 3.58e-05	🟢 5.49e-05	🟡 1.88e-05	🔴 2.23e-04	🟢 2.12e-04	🟡 5.81e-05
HGBM	🟡 0.003699	🟢 3.52e-05	🟡 2.57e-05	🟢 3.17e-05	🟢 3.47e-05	🟢 3.75e-05	🟡 1.86e-05	🟡 2.97e-04	🟢 3.08e-04	🟡 8.41e-05
XGBM	🟡 0.003714	🟢 6.81e-05	🟢 9.07e-05	🟢 1.24e-04	🟢 8.34e-05	🟡 1.59e-04	🟢 5.98e-05	🔴 7.82e-04	🟢 6.77e-04	🔴 3.14e-04
LGBM	🟡 0.003829	🟢 3.69e-05	🟡 2.66e-05	🟢 3.26e-05	🟢 3.54e-05	🟢 3.57e-05	🟡 1.98e-05	🟡 2.62e-04	🟢 2.76e-04	🟡 6.78e-05

Quick read for Table 3:

• strongest amplitude rows are now clearly ERT, HGBM, and RF;

• the hardest amplitude columns remain 156, 162, and 240;

• MLP is fully red on the amplitude side and should not be treated as a near-closure branch.

Table 4 - Phase MAE Full-Matrix Replication

Paper-side repository-owned reconstruction:

Model	1	3	39	40	78	81	156	162	240
SVM	0.002200	0.0330	0.0270	0.0610	0.1900	0.1300	1.200	0.4900	0.4900
MLP	0.007200	0.0650	0.0620	0.0800	0.1600	0.1500	1.900	0.7800	0.7000
RF	0.002000	0.0240	0.0280	0.0370	0.0740	0.0530	0.5100	0.2300	0.2500
DT	0.002100	0.0300	0.0360	0.0430	0.0900	0.0660	0.5200	0.2000	0.2300
ET	0.002400	0.0310	0.0350	0.0510	0.0940	0.0870	0.7100	0.2800	0.2600
ERT	0.002200	0.0270	0.0280	0.0400	0.0760	0.0560	0.5300	0.2000	0.2300
GBM	0.002000	0.0240	0.0300	0.0360	0.0740	0.0530	0.5400	0.2500	0.2900
HGBM	0.001900	0.0200	0.0210	0.0400	0.0910	0.0570	0.7400	0.3500	0.3600
XGBM	0.001900	0.0240	0.0320	0.0610	0.1400	0.0910	0.9600	0.5400	0.3900
LGBM	0.001800	0.0210	0.0210	0.0400	0.0950	0.0550	0.7400	0.3500	0.3400

Repository-side analogous matrix:

Model	1	3	39	40	78	81	156	162	240
SVM	🟢 0.002177	🟢 0.0323	🟢 0.0224	🟢 0.0573	🟢 0.1892	🟢 0.1230	🟢 1.088	🟡 0.5274	🟢 0.4320
MLP	🔴 0.009958	🟡 0.0802	🟡 0.0691	🟢 0.0788	🟡 0.1734	🟡 0.1655	🟢 1.562	🟡 0.9248	🟡 0.7559
RF	🟢 0.001943	🟡 0.0247	🟢 0.0268	🟢 0.0367	🟢 0.0516	🟢 0.0480	🟢 0.4251	🟡 0.2305	🟡 0.2770
DT	🟡 0.002236	🟢 0.0273	🟢 0.0323	🟡 0.0455	🟢 0.0745	🟢 0.0636	🟢 0.4900	🟡 0.2461	🟡 0.2873
ET	🟡 0.002964	🟡 0.0319	🟢 0.0328	🟢 0.0442	🟡 0.1125	🟡 0.0916	🟢 0.6079	🟡 0.3136	🟡 0.2742
ERT	🟡 0.002372	🟡 0.0281	🟡 0.0281	🟢 0.0345	🟢 0.0636	🟢 0.0492	🟢 0.3967	🟡 0.2125	🟡 0.2699
GBM	🟢 0.001883	🟢 0.0238	🟢 0.0234	🟡 0.0379	🟢 0.0608	🟢 0.0509	🟢 0.4678	🟡 0.2797	🟡 0.3021
HGBM	🟢 0.001846	🟡 0.0249	🟢 0.0204	🟢 0.0385	🟢 0.0698	🟢 0.0514	🟢 0.6100	🟡 0.3627	🟡 0.4136
XGBM	🟡 0.002165	🟡 0.0269	🟢 0.0299	🟢 0.0596	🟢 0.1181	🟢 0.0864	🟢 0.8922	🟢 0.5388	🟡 0.4270
LGBM	🟡 0.001890	🟡 0.0256	🟢 0.0204	🟢 0.0372	🟢 0.0747	🟢 0.0475	🟢 0.6092	🟢 0.3491	🟡 0.3904

Quick read for Table 4:

• phase MAE is currently the healthiest matrix of the three;

• strongest rows are RF, ERT, GBM, and HGBM;

• the hardest columns remain 162 and especially 240.

Table 5 - Phase RMSE Full-Matrix Replication

Paper-side repository-owned reconstruction:

Model	1	3	39	40	78	81	156	162	240
SVM	0.003100	0.0420	0.0440	0.0970	0.3200	0.2000	1.800	1.100	1.100
MLP	0.0130	0.0840	0.0770	0.1100	0.2400	0.2200	2.200	1.200	1.100
RF	0.002800	0.0330	0.0430	0.0550	0.1600	0.0820	1.200	0.6800	0.6300
DT	0.002800	0.0420	0.0610	0.0610	0.2000	0.1000	1.300	0.7300	0.6700
ET	0.003300	0.0460	0.0620	0.0740	0.2300	0.1500	1.500	0.9300	0.6800
ERT	0.003600	0.0400	0.0440	0.0600	0.1800	0.1100	1.200	0.6400	0.5800
GBM	0.002600	0.0340	0.0450	0.0550	0.1800	0.0840	1.300	0.7100	0.7100
HGBM	0.002500	0.0290	0.0270	0.0600	0.1900	0.0850	1.300	0.7000	0.7400
XGBM	0.002800	0.0330	0.0430	0.0890	0.2300	0.1300	1.400	0.8100	0.7600
LGBM	0.002500	0.0300	0.0280	0.0600	0.1900	0.0820	1.300	0.7000	0.7100

Repository-side analogous matrix:

Model	1	3	39	40	78	81	156	162	240
SVM	🟢 0.002908	🟢 0.0416	🟢 0.0350	🟢 0.0937	🟢 0.3139	🟢 0.1943	🟢 1.637	🟡 1.198	🟢 0.9522
MLP	🟡 0.0155	🟡 0.1022	🟡 0.0899	🟡 0.1119	🟡 0.2580	🟡 0.2296	🟢 1.912	🟡 1.484	🟡 1.252
RF	🟢 0.002667	🟡 0.0351	🟡 0.0483	🟡 0.0551	🟢 0.1250	🟢 0.0681	🟢 0.9644	🟡 0.7472	🔴 0.8462
DT	🟡 0.003114	🟢 0.0391	🟢 0.0601	🟡 0.0707	🟢 0.1513	🟢 0.0957	🟢 1.226	🟡 0.8916	🔴 0.8720
ET	🔴 0.004396	🟢 0.0429	🟢 0.0579	🟢 0.0701	🔴 0.3472	🟡 0.1658	🟢 1.220	🟡 0.9307	🟡 0.7295
ERT	🟡 0.003630	🟡 0.0409	🟡 0.0549	🟢 0.0541	🟢 0.1694	🟢 0.0784	🟢 0.9129	🟡 0.7186	🔴 0.7573
GBM	🟢 0.002510	🟡 0.0343	🟢 0.0401	🟡 0.0570	🟢 0.1399	🟢 0.0759	🟢 1.019	🟡 0.8248	🟡 0.8578
HGBM	🟡 0.002563	🟡 0.0343	🟡 0.0326	🟢 0.0593	🟢 0.1377	🟢 0.0757	🟢 1.035	🟡 0.8014	🟡 0.9239
XGBM	🟡 0.003357	🟡 0.0373	🟡 0.0456	🟢 0.0883	🟢 0.1889	🟢 0.1192	🟢 1.309	🟡 0.9495	🟡 0.9021
LGBM	🟡 0.002605	🟡 0.0365	🟡 0.0329	🟢 0.0582	🟢 0.1492	🟢 0.0686	🟢 1.054	🟡 0.8139	🔴 0.8967

Quick read for Table 5:

• phase RMSE replication is meaningfully harder than phase MAE;

• strongest rows are RF, HGBM, GBM, and ERT;

• the dominant unresolved columns remain 240 and 162.

Supporting Summary Reading Rule

For the clarified Track 1 scope, Tables 2, 3, 4, and 5 are the real paper-matching evaluation surfaces:

• Table 2: amplitude A_k MAE

• Table 3: amplitude A_k RMSE

• Table 4: phase phi_k MAE

• Table 5: phase phi_k RMSE

Table 6 remains a useful harmonic-level support summary, but it must not replace the four matrix readings above.

Supporting Harmonic-Wise Offline Result

The latest completed repository-owned harmonic-wise campaign is:

• track1_extended_overnight_campaign_2026_04_13_13_31_57

Winning validation summary:

• output/validation_checks/paper_reimplementation_rcim_harmonic_wise/2026-04-13-15-11-49__track1_hgbm_h01_wide_depth_2_campaign_run/validation_summary.yaml

Winning companion report:

• doc/reports/analysis/validation_checks/2026-04-13-15-12-35_paper_reimplementation_rcim_harmonic_wise_track1_hgbm_h01_wide_depth_2_campaign_run_harmonic_wise_comparison_report.md

Campaign results report:

• doc/reports/campaign_results/2026-04-13-16-16-23_track1_extended_overnight_campaign_results_report.md

Current best paper-faithful offline result:

• selected harmonics: 0, 1, 3, 39, 40, 78, 81, 156, 162, 240

• feature set: base_only

• validation mean percentage error: 9.830%

• test mean percentage error: 8.707%

• oracle test mean percentage error: 2.749%

• current Target A status: not_yet_met

The repository now also includes a stricter exact-paper validation branch:

• script: scripts/paper_reimplementation/rcim_ml_compensation/run_exact_paper_model_bank_validation.py

• config: config/paper_reimplementation/rcim_ml_compensation/exact_model_bank/baseline.yaml

• scope: recovered rpm, deg, tor inputs; exact ampl_k / phase_k targets; exact family bank; per-target ONNX export

• prepared campaign plan: doc/reports/campaign_plans/2026-04-10-17-04-41_exact_paper_model_bank_campaign_plan_report.md

• prepared launcher: scripts/campaigns/run_exact_paper_model_bank_campaign.ps1

• campaign results report: doc/reports/campaign_results/2026-04-10-19-54-02_exact_paper_model_bank_campaign_results_report.md

• open-cell repair campaign results report: doc/reports/campaign_results/2026-04-13-22-55-28_track1_exact_paper_open_cell_repair_campaign_results_report.md

This exact branch is now implemented, executed, and operationally stabilized. Its latest paper-closure-first campaign result confirms:

• best campaign bookkeeping run: exact_open_cell_paper_family_reference

• harmonic-level status improved from 7 partial / 3 open to 8 partial / 2 open

• no new numeric paper-target cells were closed

Its promoted full-bank structural reference run is:

• exact_full_bank_strict_reference

with:

• winning family RF

• winner mean component MAPE 18.369%

• 200 exported ONNX files

• 0 failed exports

Important scope boundary:

• this exact branch validates recovered-family fitting and per-target ONNX export stability;

• it is the canonical closure source for Track 1 table replication;

• the harmonic-wise TE-curve benchmark remains a supporting diagnostic branch for Target A, not the primary Track 1 completion gate.

What the second iteration established:

• the full RCIM set still outperforms all reduced harmonic subsets;

• the engineered operating-condition features did not improve the full-RCIM branch in this campaign;

• the reduced subsets 0,1,39 and 0,1,39,40 are weak final targets because even their truncation-only oracle stays above 4.7%;

• the main remaining gap is now better localized to predictor design, especially the dominant h0 term and a smaller late-harmonic cluster.

What the extended overnight campaign added:

• the shared offline evaluator now has a new promoted winner: track1_hgbm_h01_wide_depth_2;

• the best harmonic-wise result improved from 8.774% to 8.707%;

• the strongest companion direction in the same batch is track1_hgbm_h01_h162240_joint_balanced at 8.720%;

• the strongest isolated late-repair direction is track1_hgbm_h81156162240_cluster at 8.778%;

• the heavy low-order escalation did not beat the lighter wide winner;

• RandomForest and engineered-feature retries still did not justify promotion over the best no-engineering HGBM variants.

Immediate next repository step:

• keep the full RCIM harmonic set as the mainline Track 1 target;

• use the exact-paper report to define the open-cell repair queue first;

• use the harmonic-wise branch only to support that repair queue, especially around:

  • track1_hgbm_h01_wide_depth_2;

  • track1_hgbm_h01_h162240_joint_balanced;

  • track1_hgbm_h81156162240_cluster;

• if the next exact-paper rerun still leaves the same cells open, move the next research step to a new target-parameterization implementation rather than another winner-centric tuning cycle.

Important interpretation:

• the repository now has both a completed extended harmonic-wise Track 1 campaign and a completed exact-paper family-bank stabilization campaign;

• the best harmonic-wise result improved from 9.403% to 8.707%, so the branch is moving in the right direction;

• the paper threshold of 4.7% remains substantially unmet, so the repository is still only partially aligned with the paper offline.

Current Comparison Verdict

Comparison Axis	Current Repository Status	Verdict
Offline winner family direction	Supporting-only harmonic direction evidence remains available, but it is not the canonical Track 1 closure surface	supporting_only
Track 1 table replication	Tables 2-5 now serialized canonically; Table 6 remains the harmonic closure summary; 1/10 harmonics fully closed and 8/10 partially closed	comparable_but_not_yet_matching
Supporting harmonic-wise TE metric	Held-out mean percentage error now available at 8.707%, still above the paper threshold 4.7%	supporting_only_not_yet_matching
Online compensation benchmark	missing	not yet comparable
End-to-end paper replication	missing	not yet comparable

Online Compensation Tracking

This section is intentionally prepared now and must be updated as soon as the repository implements online compensation tests.

Repository Online Results

• Status: not yet available

• Required future fields:

  • robot profile uncompensated TE RMS and TE max;

  • robot profile compensated TE RMS and TE max;

  • robot profile reduction percentages;

  • cycloidal profile uncompensated TE RMS and TE max;

  • cycloidal profile compensated TE RMS and TE max;

  • cycloidal profile reduction percentages;

  • exact harmonic set used in each online test;

  • execution path used for the test.

Online Comparison Rule

Once repository-owned online compensation tests exist, update both:

• this report;

• doc/reports/analysis/Training Results Master Summary.md

At that point the project can present a real paper vs repository end-to-end comparison instead of the current offline-only comparison.

Missing Pipeline For A Real Table 9 Comparison

The exact missing pipeline is:

1. a repository-owned harmonic-wise prediction workflow that outputs the same practical quantities used in the paper, namely amplitude and phase terms for selected harmonics across operating conditions;

2. a TE reconstruction workflow from those predicted harmonic components;

3. a motion-profile playback workflow for the Robot and Cycloidal style profiles used as the final benchmark;

4. an online compensation loop that applies the reconstructed TE correction during motion execution rather than only offline evaluation;

5. a measurement and reporting path that records uncompensated versus compensated TE RMS and TE max in a Table 9 style format;

6. a repository-owned final comparison report that states whether Target A and Target B were met.

Until those six pieces exist, the repository results remain strong offline training results, but not yet a true reproduction of the paper benchmark.

Implementation Priority

Implement Now

• harmonic-wise prediction of A_k and phi_k

• TE reconstruction from the predicted harmonic terms

• offline motion-profile playback for Robot and Cycloidal style profiles

• paper-comparable offline validation protocol to close Target A

• repository-owned shared offline evaluator for direct-TE model families under the same final TE-curve percentage-error protocol

• evaluation of current best direct-TE families under that shared evaluator

• dual-track reporting that keeps paper-faithful and direct-TE result-level comparisons separate

These four items belong to the immediate repository branch because they create the stable offline baseline that the online branch will later depend on.

This immediate branch should now be read as an explicit intermediate stage between completed Wave 1 and the later Wave 2 temporal-model work.

Implement Later

• online compensation loop execution in the future TestRig / online branch

• uncompensated vs compensated TE RMS and TE max measurement

• final Table 9 style benchmark report to close Target B

These items should be treated as the follow-up online branch, not as the first implementation step, because they only become trustworthy once the offline harmonic prediction and reconstruction stack is already stable.

The future Wave 2 temporal-model branch also stays in the roadmap, but it is no longer the immediate next branch. It should open only after the harmonic-wise comparison framework is implemented and reviewed.

Sources

• reference/RCIM_ML-compensation.pdf

• doc/reference_summaries/03_RCIM_ML_Compensation_Project_Summary.md

• doc/reports/analysis/Training Results Master Summary.md

• output/registries/program/current_best_solution.yaml